Arthur C.

Detection and classification of non stand-replacing vegetation disturbances to understand and qualify forests resilience to climate change in Europe.

Europe, as a union of nations, is focused on achieving its climate goals to protect the biosphere. European forests, which cover more than 40% of the land, play several roles in climate change mitigation, a) they remain a natural and safe habitat for many species and therefore help to maintain biodiversity and b) they represent the largest terrestrial greenhouse gas emissions (GHG) sink absorbing the equivalent of 7% of the EU's total GHG every year. Nevertheless, those forests are threathened by human activities (tourism, limbering, etc.) and natural hazards (droughts, fire, insect attack, etc.). Those events can considerably alter biomass and play an important role in the forests' evolution. Indeed, the forest resilience to non-stand replacing events (droughts, insects attacks) is not completely understood due to a lack of accurate monitoring data on biomass before and after the events. The aim of this thesis is then to leverage the powers of satellite imagery and machine learning to accurately study these events and their impact on biomass evolution.

Joined Kayrros as a data scientist in the Biomass Team.

I developed :

- several deep learning models able to segment forests on satellite imagery and this at several spatial resolutions and on all the region of the world. The training was also performed to provide spatio-temporal stability during the inference in order to produce accurate and stable time series of segmentation.

- several models of change points detection based on the Pelt algorithm to segment the cycle stages of plantation from RADAR signals.

Multi-objective optimization of CO2 emissions, comfort, and cost of buildings via a genetic algorithm:

- creation of an encoding system to represent variables of interest as a sequence of integers. This representation has the particularity of being multi-ploid: like in DNA, several alleles compete to describe the building's phenotype (wall systems and materials used for each layer).

- adaptation of the NSGA-II algorithm proposed by Deb to work with multi-ploidy using a discrete random variable to select the dominant allele.

- study of convergence and parameter tuning of the algorithm through the creation of metrics.

- development of a comfort metamodel using a Gaussian process (Kriging) to accelerate calculation time by a factor of 10.

- validation of the metamodel's performance for use instead of EnergyPlus.

- visualization of the best building versions (phenotype: wall systems and materials, objectives: CO2, comfort, cost) with parallel coordinates.